关键词: 股价波动率, 投资者情绪, ChatGPT, 空洞时间卷积, 图注意力

Abstract: Stock price volatility stands as a pivotal research focus within the financial sector, drawing scholars into the intricate dynamics that underpin stock market movements. Recognizing the complexities involved in forecasting stock price volatility, this research embarks on an extensive analysis using daily data spanning from January 4, 2010 to September 22, 2023. For empirical rigor, six prominent Chinese publicly traded companies—namely, Industrial and Commercial Bank of China, Kweichow Moutai, GD Power Development, Daqin Railway, Yangtze Power, and East China Pharmaceutical—are meticulously selected as the study’s focal samples. Leveraging the capabilities of fine-tuned large-scale language models, this study ventures into the realm of sentiment analysis, delving deep into investor commentary on stocks. Through this, an investor sentiment index is meticulously constructed. Complementing this, the research introduces an avant-garde stock price volatility prediction model that amalgamates graph attention mechanisms with dilated temporal convolutional networks. This innovative approach is designed to forecast stock price volatility, anchoring its predictions on a nonlinear and non-stationary indicator framework.
Navigating through this intricate landscape requires an integrative methodological approach. A salient feature is the employment of fine-tuned large-scale language models for a nuanced sentiment analysis of investor-generated stock commentary. The insights garnered from this analysis culminates in the development of an investor sentiment index, providing a quantifiable measure of market sentiment. In tandem, the research unveils the GAT-TCN model, an innovative fusion of graph attention mechanisms and dilated temporal convolutional networks. This model is meticulously crafted to enhance the precision of stock price volatility predictions, operating within a nonlinear and non-stationary indicator milieu.
The empirical findings of this study are significant. Firstly, the superior efficacy of fine-tuned large-scale language models in financial sentiment classification tasks is unequivocally established. Secondly, the integration of the investor sentiment index into the predictive framework markedly elevates the accuracy and reliability of stock price volatility forecasts. Notably, the GAT-TCN network emerges as a frontrunner, showcasing enhanced predictive prowess compared to established models such as BiLSTM, CNN-LSTM, and SVM. The ramifications of these findings are profound, offering actionable insights for harnessing the potential of fine-tuned language models in financial sentiment analysis and advocating for the broader adoption of the GAT-TCN framework within the financial ecosystem. Such advancements hold transformative implications, paving the way for more precise stock price predictions and fostering interdisciplinary collaborations.
Although this study has significantly advanced our understanding of how investor sentiment predicts stock price volatility, it also emphasizes that the GAT-TCN model exhibits superior performance compared to other models that have been predicted. However, several limitations also point to possible avenues for future research. This study employs the GAT-TCN model, which has demonstrated high forecasting accuracy, but future research can continue to optimize the structure of the forecasting model as technology advances. Secondly, the findings presented in this study are based on the text of six stocks, which, although representative and broad, may not capture the entire sample of stocks. Consequently, the findings may be limited and the generalisability of the findings will be enhanced in the future through the inclusion of a larger sample of stocks.

Key words: stock price volatility, investor sentiment, ChatGPT, dilated temporal convolution, graph attention